Coloring Book
Generator
Unpaired style
transfer using
CycleGAN
-Aanchal Sahu
-Venkatakrishnan
Ranganathan
1
Why should you care as adults?
2
Offline stress reduction alternative
Shortcut
Encourages artistic expression and
motor skills
Can call it a hobby
Sense of achievement
FiftyOne
FiftyOne provides
powerful tools for
visualizing and
exploring data.
The combined power of
CLIP and FiftyOne
allows for accurate and
efficient image
similarity search and
retrieval.
3
CycleGAN
Unpaired
Image to
Image style
transfer
Capturing
common
characteris
tics of a
set of
images
Applying
that to
another set
of images.
Two
Generators
and two
discriminat
ors
4
Design
Least Squares Generative Adversarial Networks
(LSGANs)
+Improved Image Quality
+Enhanced Stability in Learning
Domain X: Real World images of Objects, animals,
people and scenery
Domain Y: Sketches and coloring book images (Black
and white)
5
Dataset Domain X
Open Images Dataset
6
Dataset Domain Y
Imagenet-Sketch Drawings from Museums
7
Experiments
Images get sharper with
more data
8
Results
9
Results
10
Evaluation: PSNR (Peak Signal-to-Noise Ratio)
Distortion, noise, information loss, and pixel intensity.
Higher PSNR = better image similarity and quality. Anything
between 30-50 dB is considered good.
11
Analysis: PSNR
The CLIP version
performs better in
the PSNR
evaluation metric.
Our scores always
lie between 27.01-
31.57 dB since the
original image and
coloring book
version look quite
different.
12
Evaluation: SSIM (Structural Similarity Index)
Structural information, luminance, contrast, and
similarity of patterns in the images.
Higher SSIM values = greater similarity and
better quality. Ranges from 0 to 1, with 1
representing a perfect match.
SSIM is more aligned with human perception than
PSNR and considers factors like brightness and
texture.
13
Analysis: SSIM
14
Wide range.
Some images
are very
similar to the
original image
while some are
very
different.
The third
attempt seems
to perform the
best for this
evaluation
metric.
Problems and Future Work
15